The Artificial Intelligence Show

The episode covers 15 questions in 57 minutes and occasionally surfaces genuine practitioner thinking - particularly on agent security, the experienced-manager thesis, and the AI confidence gap - but the majority of responses are padded with hedging, self-anecdotes, and vague generalizations. Insight-per-minute ratio is low for the runtime.

“the confidence level I have that the safeguards we're putting in place will hold are lower by the day. They're not going in the opposite direction. There's a reason the government forced Anthropic to withhold their smartest models. It's because they can't control them.”

“I don't think most organizations are at that level of sophistication, but the ones who are asking these questions and the even more advanced ones are saying these models out of China, they're not so bad, they're 10x cheaper than the US models.”

Most frameworks - AI councils, governance layers, human-machine symbiosis - are standard AI thought-leadership content. The candid skepticism about Anthropic's walled-garden claims and the argument that experienced managers will outpace middle managers due to AI are the genuinely fresher takes, but the rest recycles common advice.

“you give Claude access to your Gmail and your calendar. There's a whole lot of stuff that happens in your Gmail that maybe other people in the company shouldn't read... Again, may 100% be true. You are taking a leap of faith that that is accurate.”

“the manager often doesn't have that advanced experience expertise you develop from failing sometimes and having enough campaigns you ran that didn't work... They don't have the full taste and judgment to know what's good and what's not good.”

Paul is a credible AI educator with real organizational deployment experience and plausible practitioner instincts, but the episode features no external guests - it is solely the two hosts - and Paul's primary domain is AI education and marketing rather than operating AI at enterprise scale in a non-AI-native business.

“I just actually approved. Two days ago I had a meeting with our COO and our Director of Operations to go through our updated AI policy.”

“we have trained this thing. It has this massive knowledge base, is trained on hundreds of FAQs. But like someone on our team actually looks at those engagements”

The episode contains scattered concrete details - 50 agent iterations at SaaStr, the Coinbase org-flattening reference, Claude's UK-summer laziness study, the $20 - 25/month licensing figure - but most claims rest on anecdote or vague quantifiers like 'most companies' and 'less than 5%' without sourcing, and no hard business metrics are ever cited.

“they had like 50 iterations of this thing before it became reliable enough to actually work in a trusted way”

“there was this weird study back in like early 2025 I think it was, that showed that Claude got lazy during the summer months in the UK”

The questions are pre-curated from class attendees and are genuinely strong for the format, but Kathy functions almost entirely as a question-reader and light affirmer rather than a skilled interviewer - she rarely follows up with challenge or precision, and Paul's claims go consistently unchallenged even when he hedges openly or speculates.

“How would you recommend folks know that a legal team or an outside IT team could help that they're the right people to analyze such things?”

“Maybe we need to build a co CEO GPT so we can put in here what would A footwear Tracy or COO be asking us”